World Bank AI
Facial-recognition attendance and data validation at institutional scale, engineered for accuracy and privacy.
About World Bank AI
DSME built a computer-vision and AI-automation platform for facial recognition, attendance tracking, and data validation, delivered for a major global institution. The system replaces manual roll-calls and error-prone paper records with a real-time, high-accuracy pipeline deployed across multiple countries and thousands of staff.
Industry · Computer Vision / Public Sector
The problem we set out to solve
Large public-sector institutions operating across borders face a deceptively hard problem: verifying who is present, where, and when, consistently and defensibly across offices with different lighting, hardware, and connectivity. Manual attendance is slow, easy to game, and produces data that cannot be trusted for payroll, security, or reporting. Any automated alternative must clear a high bar for recognition accuracy across diverse populations while respecting strict privacy and data-governance expectations. It also has to run reliably in facilities where cloud connectivity is intermittent, without shipping raw biometric imagery off-site. The engineering task was to hit near-perfect accuracy and airtight data handling at the same time, at scale, in production.
How we built it
Data collection and model training
We assembled representative enrollment datasets and trained facial-recognition models in TensorFlow, tuning for accuracy across varied demographics, poses, and lighting conditions.
Edge and on-prem inference
Recognition runs close to the capture point using OpenCV pipelines and containerized services, so sensitive imagery is processed locally rather than streamed to a central cloud.
Privacy-first data validation pipeline
A dedicated validation layer cross-checks matches, flags anomalies, and stores only the minimum derived data needed, keeping the system aligned with strict governance requirements.
Cloud training and MLOps on AWS
Model training, versioning, and evaluation are orchestrated on AWS SageMaker, with Docker and FastAPI packaging inference services for consistent, repeatable multi-country rollout.
Key capabilities
High-accuracy facial recognition
Deep-learning models identify enrolled staff in real time under real-world conditions.
Automated attendance tracking
Presence is captured, timestamped, and logged without manual sign-in or supervisor intervention.
Data validation engine
Every match is cross-verified and anomalies are surfaced before records enter downstream systems.
Edge-ready deployment
Containerized inference runs on local hardware for low latency and data residency.
FastAPI service layer
Recognition and validation are exposed through clean, high-throughput APIs for integration.
Multi-country rollout tooling
SageMaker-trained models and Docker images standardize deployment across every site.
Outcomes that moved the needle
The platform reached 99.2% recognition accuracy in production while tracking more than 5,000 staff daily, and was rolled out successfully across 3 countries. Automated, validated attendance data replaced manual processes, giving the institution a trustworthy record for security and reporting without compromising on privacy or data governance.
